population management in clouds is a do-it-yourself technology

.

Questions and Objectives• (1a) How to steer the global network behavior resulting fromself-organization where agents have their own goals and actions? -- globalvs local/self optimization

• (1b) How to build and leverage the knowledge produced at the micro (agent) andmacro (network) levels? How to correlate and make sense out ofmeasurements and data from heterogeneous sources? -- e2eperformance heterogeneity

• (2) How to make the operations scale? To which extent are the newtechniques and solutions actionable, "computable"...? -- complexity problem

• (3) How can the management techniques leverage on the new properties andfeatures introduced by adaptive and cooperative agents, i.e. applying thesame techniques to the management plane itself. -- heterogeneityproblem

M.Zhanikeev -- [email protected] Population Management in Clouds is a Do-It-Yourself Technology -- bit.do/*** 2/25...

2/25

.

This Talk is about Cloud Networks• a network of DCs is a cloud -- say, Amazon EC2 with 9 regions

• a network of clouds is a federation -- say, Amazon x 1dc + IDCF x20dcs in Japan◦ we already start hitting the heterogeneity wall, since there is no solution yet, no

such federations exist today, but some are planned 03

• .. moving further to network edge, we find Akamai with its 30k+ cloudlets02 -- a class of its own◦ heterogeneity problem? in hardware NO, but in e2e network performance STRONG YES◦ but, Akamai has all the control, so, they manage to balance the load properly

• ... now, having incorporated all the devices at network edge you have thefog cloud

• ... on a side note... why cloud networks? ... well, because most networkingsoon will be done inside and/or by clouds

03 "Chameleon project" https://www.chameleoncloud.org (2014)

02 B.Frank+8 "Pushing CDN-ISP Collaboration to the Limit" ACM SIGCOMM Computer Communication Review (2013)


3/25

.

Cloudification Timeline and Scope

WiFi

Sca

le/s

ize o

f clo

udifi

edite

ms

Timeline

Mid-2015

Probablyshould not

cloudify

Should definitely cloudify

DataCenters

Racks IoV > VC

Desktop

APs,Beacons

Smartphone

Sensor

IoTDelegatednetworking


4/25

.

One Cloud Service.VNE Problem......stands for Virtual Network Embedding

APP

Cloud/DC

APP

APP

VM Container

APP

Cloud/DC

APP

APP …

• service can provide thegraph of its apps

• graph can include end users inform of DC-user e2enetwork performance

• federated cloud known itsactual topology and can useVNE to map 04

04 M.Zhanikeev "A New VNE Method for More Responsive Networking in Many-to-Many Groups" 7th ICUFN (2015)


5/25

.

Many Cloud Services.VNE......is the same, no difference

Cloud A Cloud B

DC DC

DC

Virtual Physical

Many

Service topology

• for many services, you mightuse a more detailedphysical topology(rack-level?)

• otherwise, the same mappingmethod

• many-service case is alreadyconsidered in literature 04

04 M.Zhanikeev "A New VNE Method for More Responsive Networking in Many-to-Many Groups" 7th ICUFN (2015)


6/25

.

OSPF, QoS Routing, VNE → complexity

• OSPF optimizes one parameter (cost/weight) for one physical graph, and is NPHard

• QoS Routing optimizes multiple parameters for one path, and is NP Hard• VNE maps one virtual graph on top of a physical graph and is NP Hard-er?• ... so, what happens when the global NOC tries to optimize many VNEs? NPHard-er-er-er?


7/25

.

Solution to the Complexity Problem

.The Simple Solution is.....

.... Do-it-Yourself, that is, to let services self-optimize• however....

◦ it is one thing to self optimize in a non-cooperative environment◦ ... and quite another when clouds offer tools/envs/platforms thathelp self-optimize better


8/25

.

Non-Cooperative Way


9/25

.

Step 1: Probe your Cloud• visualization of active probingacross 8 Amazon regions 05

california

ireland

oregon

saopaulo

singapore

sydneytokyo virginia

same DClarge bulk transfer

california

ireland

oregon

saopaulo

singapore

sydney

virginia

DC - other DCe2e delay

05 M.Zhanikeev "Performance Management of Cloud Populations via Cloud Probing" IPSJ JIP (related to http://goo.gl/f3QpiW) (2015)


10/25

.

Step 2: Optimize your service network

• change DCs for some members of your population, gradually optimize 05

Pooler model

0 100 200 300 400 500Ordered list of values

0

20

40

60

80

100

Com

plet

ion

time

(s)

Do nothingOptimize



11/25

.

Cooperative Clouds


12/25

.

Traditional vs Akamai vs LHAP

Local hardware

Virtual Machine

Cloud Platform (Xen)

App App App …

Multitenant

environment

Physical Machine (PM)

Tools

API Layer Control Layer

Local Hardware Awareness Platform (LHAP)

…

Virtual Machine

App

Tools …

Virtual Machine

App Virtual Machine

App App App …

Global network (to Internet)

In-box network

Fig1a: Conventional Platform

…

MiniOS

App

SSD

Operating System

Virtual Machine

App App App …

Multitenancy impossible


Fig1b: Mini Cache


Operating System


HDD Local Network



Fig1c: LHAP (proposed)

Cloud Platform (Xen) Operating System

DtL: Direct to Local

Local hardware

SSD HDD


06 M.Zhanikeev "A Cloud Visitation Platform to Facilitate Cloud Federation and Fog Computing" IEEE Computer (2015)


13/25

.

Akamai MiniCache Technology

Local hardware


Multitenant

environment


…

Virtual Machine

App Virtual Machine

App App App …


Fig1a: Conventional Platform

…

MiniOS

App

SSD

Operating System

Virtual Machine

App App App …

Multitenancy impossible


Fig1b: Mini Cache


Operating System


HDD

• depends on MiniOS, part of Xen• MiniCache allows MiniOS-based appsto use SSD/HDD connected to that PM

• MiniCache (storage) and ClickOS (network)are similar technologies, the letter is usedfor SDN

• demerits

1. restricted to PM2. storage only3. no concurrency control or load balancing



14/25

.

Local Hardware Awareness (LHA)

Virtual Machine

App App App … Tools

API Layer Control Layer

Local Hardware Awareness Platform (LHAP)

…

Virtual Machine

App

Tools

In-box network

Local Network



Fig1c: LHAP (proposed)

Cloud Platform (Xen) Operating System

DtL: Direct to Local

Local hardware

SSD HDD

• not restricted to resource type, anyresource◦ storage, sensors, Hadoop/MapReduce jobs, activeprobing, ...

• not resricted to same PM, API Layer knowsthe resources offered for discovery

• Tools in VMs (and containers) interface withLHAP

• LHAP is optional -- legacy VMs will not knowit's there



15/25

.

Example Application : CDN


16/25

.

CDN on Federated Clouds

Large Cloud

Small Cloud

Small Cloud

CDN Manager

CDN

Local clients


17/25

.

Cloud Streaming in the Wild

VMpopulation

CurrentSources

ServiceProvider

(SP)

Tracker

ServiceProvider

(SP)

Parentpeers

P2Pstreaming

Cloudstreaming

ClientClient

07 M.Zhanikeev "Multi-Source Stream Aggregation in the Cloud" ISBN 978-1-118-57521-5, Wiley (2014)


18/25

.

Caching options with LHAP• (don't laugh but) caching methods have been lost on cloud-based CDN -- a trueclean slate

• yet, LHAP makes it possible to add additional caches• note: Cache 1 and Cache 2 are completely different in nature -- VMsare short-lived

CDN

Main

many

Manager

Storage Large

Federation, fog clouds VM

Cache 1 Cache 2

Manage content

CDN Traffic (End Users)

many

Cloud


19/25

.

Back to Networks: Link with NGN


20/25

.

Key Viewpoints

1. traditional NGN in a recent book 08

◦ virtualization is key (SDN, NFV), among other reasons for supporting mobility

2. my own viewpoint is different◦ because of hotspot traffic (bulk transfer = BigData = hotspots), e2e circuits

(cut-through) are key 09◦ circuits are made in a distributed/social/sensing manner 10 close to the Y.mnm 11

story◦ as a guy who was working on a distributed network of active probes 12 (now cloud

boxes), the Y.mnm story is close to mine

08 T.Janevski "NGN Architectures, Protocols and Services" Wiley (2014)

09 M.Zhanikeev "The Next Generation of Networks is all about Hotspot Distributions and Cut-Through Circuits" http://goo.gl/a16JtD (2015)

10 M.Zhankeev "A City Traffic Model for Optical Circuit Switching in Data Centers" http://goo.gl/5z2V6F (2015)

11 "Management of Performance Measurement for NGN" ITU-T Y.2173, before that Y.mnm (2008)

12 M.Zhanikeev "A Home Gateway Box with Meter, Probe and L2 QoS Policy Edge" IEEE COMPSAC (also http://goo.gl/9graXy) (2013)


21/25

.

The Y.mnm Story : Active Probes• 2 parts: active probing anddistributed management11

• between OSPF and QoSRouting, closer to thelatter, but that was beforeVNE was defined

• active probing is its ownworld of research 13

• among smaller projects(smartphones, etc.)Google-backed M-Lab is thebiggest

11 "Management of Performance Measurement for NGN" ITU-T Y.2173, before that Y.mnm (2008)

13 Y.Tanaka, M.Zhanikeev "Active Network Measurement: Theory, Methods, and Tools" ITU Association of Japan, ISBN 978-4916128072 (2009)


22/25

.

IETF Liason with Y.mnm


23/25

.

Wrapup

• complexity wise, VNE formulation is close to reality, but is NPHard-er-er-er◦ effective heuristics are unlikely

• theDo-it-Yourself (DiY) Approach can resolve thecomplexity problem◦ non-cooperative form: networks optimize themselves (example: Cloud Probing

05◦ assisted form: clouds/networks can implement a Local AwarenessFeature (like LHAP 06) to improve self-optimization

• the Assisted DiY also leads to better coordination across competingservice networks




24/25

.

That’s all, thank you ...


25/25